|
|
--- |
|
|
language: multilingual |
|
|
tags: |
|
|
- classification |
|
|
- emotions |
|
|
license: apache-2.0 |
|
|
metrics: |
|
|
- precision |
|
|
- recall |
|
|
- f1-score |
|
|
- accuracy |
|
|
--- |
|
|
|
|
|
# Emotion Classification Model |
|
|
|
|
|
## Model Description |
|
|
This model is a fine-tuned version of `xlm-roberta-large` for multilingual emotion classification tasks. It is trained to classify text into 9 distinct emotion categories: |
|
|
|
|
|
- **Anger (0)** |
|
|
- **Fear (1)** |
|
|
- **Disgust (2)** |
|
|
- **Sadness (3)** |
|
|
- **Joy (4)** |
|
|
- **Enthusiasm (5)** |
|
|
- **Hope (6)** |
|
|
- **Pride (7)** |
|
|
- **No emotion (8)** |
|
|
|
|
|
The model is designed to analyze input text and predict the corresponding emotion, including the neutral "No emotion" category. |
|
|
|
|
|
--- |
|
|
|
|
|
## Model Performance |
|
|
The model was evaluated on a dataset of 12,022 examples (10% of all data). Below is a summary of the performance metrics across all categories: |
|
|
|
|
|
| Emotion | Precision | Recall | F1-Score | Support | |
|
|
|----------------|-----------|--------|----------|---------| |
|
|
| Anger (0) | 0.70 | 0.50 | 0.59 | 2936 | |
|
|
| Fear (1) | 0.56 | 0.13 | 0.21 | 317 | |
|
|
| Disgust (2) | 0.56 | 0.35 | 0.43 | 105 | |
|
|
| Sadness (3) | 0.69 | 0.40 | 0.51 | 334 | |
|
|
| Joy (4) | 0.58 | 0.56 | 0.57 | 427 | |
|
|
| Enthusiasm (5) | 0.42 | 0.15 | 0.23 | 544 | |
|
|
| Hope (6) | 0.50 | 0.20 | 0.29 | 777 | |
|
|
| Pride (7) | 0.57 | 0.32 | 0.41 | 354 | |
|
|
| No emotion (8) | 0.64 | 0.88 | 0.74 | 6228 | |
|
|
|
|
|
### Overall Metrics |
|
|
- **Accuracy**: 64% |
|
|
- **Macro Average**: Precision: 0.58, Recall: 0.39, F1-Score: 0.44 |
|
|
- **Weighted Average**: Precision: 0.63, Recall: 0.64, F1-Score: 0.61 |
|
|
|
|
|
--- |
|
|
|
|
|
## Usage |
|
|
|
|
|
### Input |
|
|
The model expects a text input in UTF-8 format. The input can be a sentence, paragraph, or any textual data. |
|
|
|
|
|
### Output |
|
|
The model outputs a predicted emotion label from the predefined categories, along with the associated confidence scores. |
|
|
|
|
|
### Example |
|
|
```python |
|
|
from transformers import pipeline |
|
|
|
|
|
classifier = pipeline("text-classification", model="uvegesistvan/wildmann_german_proposal_0") |
|
|
|
|
|
text = "Ich bin so glücklich über die Fortschritte, die ich gemacht habe!" |
|
|
prediction = classifier(text) |
|
|
|
|
|
print(prediction) |
|
|
# Output: [{'label': 'Joy', 'score': 0.85}] |
|
|
``` |
|
|
|
|
|
## Training Data |
|
|
The model was trained on a dataset containing labeled examples for 9 emotions. All training data was on german. The "No emotion" category is the most represented in the dataset. |
|
|
|
|
|
## Limitations and Bias |
|
|
- The model's performance may vary across languages or cultural contexts not well-represented in the training data. |
|
|
- The "Fear" and "Enthusiasm" categories have lower recall and F1 scores, indicating potential underperformance in these classes. |
|
|
|
|
|
|
